Wages and Health Worker Retention: Evidence

from Public Sector Wage Reforms in Ghana

James AntwiGreenwich School of Management

University of Wales

David Phillips 1

Department of EconomicsGeorgetown University

February 2011

1Corresponding author; can be contacted at dcp34@georgetown.edu. We have val-ued feedback from Michael Clemens, Garance Genicot, Billy Jack, Arik Levinson, AnnaMaria Mayda, and seminar participants at the 2010 NEUDC Conference, the George-town Public Policy Institute Development Workshop and the Georgetown GraduateStudent Seminar. We would also like to thank Dr. Ebenezer Appiah-Denkyira, Mar-garet Chebere, Victor Ekey, and staff at the Ministry of Health, Ghana Health Service,Nurses and Midwives Council, and CAGD that assisted by providing data and insight.We are also grateful to Chris Herbst, Karima Saleh, Agnes Soucat, and others in theWorld Bank (AFTHE) for their support. During this research, Mr. Phillips workedfor the World Bank (AFTHE) as a consultant and Mr. Antwi worked for the GhanaMinistry of Health. The opinions expressed herein do not necessarily reflect those ofthe World Bank or the Government of Ghana. All errors and omissions remain theresponsibility of the authors.

Abstract

Can governments in developing countries retain skilled health workers by raisingpublic sector wages? We investigate this question by way of a policy-inducednatural experiment, in which the Government of Ghana restructured the payscale for health workers employed by the government. We find that a ten per-cent increase in wages decreases annual attrition from the public payroll by 0.9to 1.6 percentage points (from a mean of about 8 percentage points) among 20-40 year-old workers from professions that tend to migrate. We find that theseeffects are concentrated among young workers, and we detect no effects amongcategories of workers that do not tend to migrate. Given Ghana’s context asa major source of skilled health professional migrants and high correlation ofour attrition measure with aggregate migration, we interpret these results asevidence that wage increases in Ghana improved retention mainly through re-ducing international migration.JEL Codes: O15, F22, J45, J61, and H51.

1 Introduction

High attrition of skilled employees can generate under-staffing in the publichealth care systems of developing countries. Wage differentials between domes-tic public employment and other options are one factor that could be drivingdoctors, nurses, and other skilled health workers to leave the public health sec-tor, often for jobs in high-income countries. Knowing the impact of wages onattrition takes on particular importance when responding to the migration ofhealth workers from low and middle-income to high-income countries. Thoughthe appropriate response to this trend is oft-debated, many developing countrieshave decided to take policy positions discouraging such migration. Increasedsalaries represent one possible but expensive option for improving retention.Knowing the impact of increased salaries on migration provides vital informa-tion to policy makers considering the costs and benefits of raising salaries asopposed to mandatory public service, improved facilities, shifting from higherto lower skilled health workers, and other options.

Despite the importance of this issue to policy makers and the centrality ofwages in the basic economic model of migration, little strong evidence exists onthe causal impact of home country wages on attrition of health workers. Thispaper aims to isolate the causal effect of wages on attrition by use of a naturalexperiment. In the ideal econometric situation, wages would be set randomly sothat any correlation between wages and attrition would be causal. Lacking thissituation, we exploit a sudden change in the compensation of different groups ofpublic sector health workers in Ghana. As a result of this policy change, healthworkers of different categories and different levels of experience received widelyvarying pay increases. Paired with administrative data, we use this naturalexperiment and a difference-in-differences strategy to test whether the groupsof health workers who received the largest raises had their attrition rates fallthe most.

We find that, on average, higher wages do decrease attrition of health workersfrom Ghana’s public payroll. A ten percent increase in wages decreases annualattrition by 0.9 to 1.6 percentage points (10 to 20 percent of the mean) among 20-40 year-old potential migrants. We find evidence that this effect is concentratedamong the youngest workers (20-35 years old) and no evidence that wages affectattrition for mid-career potential migrant workers (35-55 years old). Amongcategories of workers that rarely migrate, we find no evidence of wages reducingattrition for early-to-mid career workers. During this time period, Ghana wasa major source of high-skilled health worker migrants, and measures of healthworker migration correlate strongly with our measure of attrition at aggregatelevels. Thus, we interpret the results as most plausibly capturing the effect ofwages in reducing international migration. Finally, we also detect an impactof wage increases in reducing attrition among late-career workers from bothpotential migrant and non-potential migrant categories. We attribute this effectto reducing attrition from retirement.

Our main results confirm the predictions of simple models of migration andemployee retention. In the simplest migration models, an increase in home

1

country wages unambiguously reduces the probability that an individual willmigrate by reducing the ‘push’ effect of a large wage gap (Borjas, 1987). Simi-larly, on-the-job search models of employee retention predict that an employerwith higher wages will retain workers for longer by competing more effectivelywith other employers (Burdett and Mortensen, 1998). While confirming themost basic models, our results contrast with simple model extensions that cancounteract or even reverse the effect of higher wages. In models of migration,home wages can encourage migration by making it possible to overcome creditconstraints that prevent financing migration costs (Lopez and Schiff, 1998). Inour context, higher wages lead to higher future pensions as well. So, in jobsearch models, higher wages could have wealth effects that encourage retire-ment or other decreases in labor supply. From the standpoint of these theories,the impact of wages on attrition is ambiguous. Our results indicate that the netimpact of wages on attrition is negative, i.e. the incentive to stay provided byhigher wages dominates such credit or labor supply effects.

The remainder of the paper is as follows: section II describes Ghana’s healthsector and the 2006 wage reforms; section III provides a short discussion of thetheory; section IV briefly reviews pertinent literature; section V describes ouridentification strategy; sections VI and VII describe the data and the results.Finally, section VIII explores some robustness checks and section IX concludes.

2 Background

2.1 Migration of Health Workers from Ghana

Ghana has long been a major source of migrants in the health sector. Likely duetheir high quality training, low wages, and English proficiency, many Ghanaianhealth workers have left for jobs abroad. Bhargava and Docquier (2008) providecross country data on physician migration into OECD countries. As shown inFigure 1, in an average year from 1991-2004 three to four percent of Ghana’sphysicians migrated annually, easily outpacing the African average. Prior to thistime period, migration rates were even higher, with Dovlo and Nyonator (2003)reporting annual migration rates of 10 to 20 percent for graduates in the 1985-1994 classes of the University of Ghana Medical School. As shown in Table 1,these migrants mainly leave for English-speaking, high-income countries. Datafrom the Ghana Nurses and Midwives Council indicate that a full 71 percent ofnurses leaving during 2002-2005 went to the UK, with most of the remainderleaving for the US. Data from Dovlo and Nyonator (2003) indicate a similarpattern for physicians.

After multiple decades of extensive migration by Ghanaian health work-ers, flows of health workers out of Ghana have slowed in recent years. Figure2 demonstrates this fact for nurses using Nurses and Midwives Council dataand administrative payroll data. Ghana’s Nurses and Midwives Council keepsstatistics on the number of requests by domestically trained nurses to have theircredentials verified for international employment. As the figure indicates, migra-

2

tion of nurses from Ghana plateaued in the early 2000’s, dropped precipitouslyin 2006, and then leveled off at a reduced rate. Attrition of nurses under age40 from the public payroll shows a similar pattern. In the same figure, attritionrates show a large drop in 2006 and subsequently stabilize, closely followingthe NMC migration data. Recent decreases in migration are also apparent forphysicians, as depicted in Figure 3. Attrition from the public sector and data onnew Ghanaian registrants to the UK’s General Medical Council show a strongcorrelation with each other as well as a drop in 2006. The close correspondencebetween migration data and attrition from the public payroll will be importantlater for the interpretation of our results. Because individual-level migrationdata is unavailable for our sample, we will use attrition from the public payrollas our dependent variable. The time-series correlation of our dependent variablewith migration measures provides an indication that attrition in our data is bestinterpreted as migration.

The recent, sudden decline in migration of health workers from Ghana begsthe question as to its causes. As Figure 4 shows, health workers migrating tothe UK can roughly double their earnings, even after adjusting for purchasingpower differences. For example, doctors in Ghana earned about 1,000 GhanaCedis per month but could earn about 2,500 Ghana Cedis per month (PPP) inthe UK. Many point to such wage gaps as the main cause of migration of skilledhealth workers to high-income countries. In 2006, at precisely the same time asthe fall in migration, the government of Ghana introduced a new wage structurefor health workers that increased earnings significantly for many health workers.While many other factors and policies in Ghana and abroad likely influencedthe decline in migration, we will focus on isolating the role that wages played.

2.2 Public Health Sector Wage Changes in Ghana

In Ghana, as in many sub-Saharan African countries, the public health systemand religious non-profit hospitals together comprise most of the health sector.In Ghana, the public health sector is mainly composed of the Ministry of Health(MOH), the Ghana Health Service (GHS), and the two main teaching hospitals:Korle Bu in Accra and Komfo Anokye in Kumasi. In addition to these facilities,some health workers are posted to military and prison hospitals. The ChristianHealth Association of Ghana (CHAG) provides most of the non-profit healthservices in the country. Together, the MOH estimates that these two sectorsemploy 90 percent of all health workers in Ghana. Most of the remainder workin the private for-profit health sector. Importantly for this study, the centralgovernment pays the salaries of both public and CHAG employees. Payrollrecords for employees of these agencies are kept at the Controller and Accoun-tant General Directorate (CAGD) in Accra. Our data will cover all workers inthis database who fall under MOH, i.e. all health workers in the country exceptfor those in the for-profit private sector and those employed at prison and mil-itary hospitals. MOH estimates that the employers in our database employ 82percent of health workers in Ghana.

In 1998, the Ministry of Health introduced the Additional Duty Hours Al-

3

lowance (ADHA) for health workers. As its name implies, the ADHA’s explicitpurpose was to compensate doctors, nurses, and other core clinical workers forunusually long hours. However, shortly after its creation in 1998, the ADHAbecame a simple salary supplement and was extended to other cadres of healthworkers. The Ministry of Health (MOH) assigned a fixed number of notionalhours to each cadre (doctor, professional nurse, etc.) of employee, and all em-ployees in the same cadre received the same number of hours. Since these hourswere paid at the worker’s usual hourly rate, the ADHA amounted to a per-centage bonus of a health worker’s base salary. Within a cadre all employeesreceived the same percentage bonus from ADHA, while different cadres receiveddifferent bonuses due to differences in notional hours assigned. This system, sup-plemented by common percentage pay raises among all employees, ensured thatthe relative pay of all healthworkers was stable from 2000-2006.

In 2006 due to budgetary pressure, the Government of Ghana desired tofold the ADHA into regular pay. Health workers also pushed for ADHA to beconverted into basic salary because the ADHA payments were often delayedand were not taken into account when determining pensions. Since ADHA hadgrown to be a very large proportion of many health workers’ pay, the governmentdecided to adopt an entirely new salary structure, the Health Salary Structure(HSS). Pay rates under the new schedule were defined based on a job evaluationthat arranged different jobs in the new salary structure according to the skillsand tasks of the job. As a result, the new salary structure gave nominal wageincreases of varying degrees to all health workers, but due to inflation and theloss of ADHA, some workers saw their real total earnings rise slowly or evendrop. Importantly, due to differences in the raises assigned to various groups ofworkers and due to the fact that some cadres benefitted more from ADHA thandid others, the new salary structure completely rearranged the relative pay ofmany workers. Finally, as a part of the arrangement made in adopting the HSS,workers’ nominal wages were frozen from 2006-2009.

Figure 5 displays these wage changes. In the wage schedule, a health workers’pay is determined by ‘grade’ and ‘step.’ Grades differentiate large promotions(principal nursing officer, medical officer, senior medical officer, specialist, etc.)while steps embody smaller promotions within a grade. In the figure, eachline represents real log wages (inclusive of ADHA) for each possible gradeXstepcombination, normalized to zero in 2003. Thus, following an individual line overtime traces the wages of a worker that is never promoted from 2003-2009. From2003-2005 the lines generally move together, demonstrating that all groups ofworkers received common percentage wage increases. But from 2005 to 2006, thewages of different groups of workers diverge. Some workers received real wageincreases of up to 50 percent while others even saw their real wages decreaseby up to 30 percent. We will exploit the variation in wages across professions,seniority, and time generated by this policy change to measure the impact ofwages on attrition.

4

3 A Simple Theoretical Framework

3.1 A Simple Model of Migration

Consider an individual choosing between continuing to work in the public healthsystem and leaving for another job. We will interpret this other option asmigrating for a job outside the country, but in principle the outside job couldbe in the private health sector or outside the health profession. Assuming alinear indirect utility function, an individual i will attrite at time t iff:

α0 + β1w∗it + δ0xit ≥ α1 + β2wit + δ1xit + cit

⇔ cit ≤ β0 + β1w∗it − β2wit + δ′xit (1)

where cit is the cost of migration, w∗it is the log wage abroad, wit is the logwage at home, and xit is a vector of individual characteristics that are valueddifferently at home and abroad (δ is the marginal value of an attribute abroadrelative to home). If F is the distribution of cit then the probability of attritionAit is:

Ait ≡ Pr [Attrition] = F [β0 + β1w∗it − β2wit + δ′xit]

In this simple model, the impact of home wages on the probability of attritionis unambiguously non-positive. Assuming F is differentiable with density f :

∂Ait

∂wit= −f (β0 + β1w

∗it − β2wit + δ′xit)× β2 ≤ 0

However, even in this model the magnitude of the impact of wages dependsgreatly on the functional form and support of F . In particular, if the wage gapis very large between the home and foreign countries, then the impact of wageswill be likely be small. Intuitively, large wage gaps move us into the ‘tail’ ofthe distribution of migration costs.1 In the case of health workers migratingfrom Ghana to their main destination in the UK, this could very well be thecase. In fact, some policy-focused research discourages salary increases as amethod for decreasing health worker migration from sub-Saharan Africa due tothe perception that salary increases will likely be ineffective due to the largewage gaps (Vujicic, et. al., 2004).

3.2 Credit Constraints

The unambiguous negative impact of home wages on attrition disappears if asimple credit constraint is added to the model. In an extreme case, suppose thata worker receives the public sector wage at time t. Then, the individual canchoose whether or not to leave, expecting that future wages will be the sameas today. Finally, suppose that the cost of migration must be financed out ofcurrent wages. Then, for an individual to migrate, they must be able to financemigration:

1Formally, as long as limw∗→∞f (·) exists, then it must be zero. As a result, the limit of∂Ait∂wit

must be zero as well.

5

cit ≤ wit (2)

Thus, an individual attrites iff (1) and (2) both hold, i.e. the probability ofattrition is

Ait = F [min{wit, β0 + β1w∗it − β2wit + δ′xit}]

For individuals with low wealth, wage increases may actually lead to highermigration rates:

∂Ait

∂wit= f (wit) ≥ 0 if wit < β0 + β1w

∗it − β2wit + δ′xit

Thus if a change in home wages reflects both an increase in current wages anda similar increase in expected future wages, the sign of the marginal effect ofhome wages on migration is an empirical question as well.

4 Literature Review

A growing literature attempts to identify and measure the determinants andimpacts of medical migration from developing countries. Much of this literaturefocuses on the impact of migration on health outcomes in the sending country(e.g. Clemens, 2007; Bhargava and Docquier, 2008). The literature on deter-minants remains largely qualitative. Two quantitative papers by Okeke (2009)on the determinants of physician migration in Africa come closest to our topic.These studies use cross-national comparisons to examine the impact of growthshocks and the introduction of the ADHA in Ghana on physician migration.We take a different approach to measuring the impact of wages on migrationby using micro data to examine the impact of a well-defined policy change onindividual decisions.

Our study is also related to the extensive literature on the determinants ofmigration, in general. The cross-country panel literature on the determinants ofmigration generally proxies home country wages with home country GDP percapita. These studies rarely detect a statistically significant impact of homecountry wages on migration. Most studies suggest that this is likely due to op-posing incentive and credit constraint effects. However, it is difficult to test thisproposition cleanly using macro data, though some studies find evidence thathigher GDP correlates with more migration from the lowest income countriesand less migration from higher income countries (e.g. Clark, et. al., 2007; Ped-erson, et. al., 2004). These studies help clarify the correlations across countriesbut often have difficulty providing results with a reasonable causal interpreta-tion; however, some studies move in this direction by using country fixed effectsand lagged GDP (Mayda; 2008). We hope to move further in this directionusing an identification based on micro-data and a natural experiment. We willalso be able to measure wages directly rather than proxying with GDP.

Studies using individual-level data attempt to measure the causal impactof wages while removing bias caused by unobserved ability and other factors.Gibson and McKenzie (2010) attempt to eliminate this bias by focusing on a

6

very narrow group of very highly-skilled individuals: valedictorians and Sci-ence Olympiad participants. This helps address endogeneity, but as they ac-knowledge, this approach necessarily limits the external validity of the results.Other studies attempt to eliminate the bias by modeling selection into migra-tion parametrically. This approach to selection, unfortunately, often requiresstrong assumptions about the parametric form of unobservables that cannot beverified.

A few papers have applied a natural experiment methodology to questionsrelated to migration and wages. Yang and Choi (2007) find that rainfall-inducedincome shocks in the Philippines cause increased migration from households thatwere previously unable to send migrants abroad. Similarly, Angelucci (2005) ar-gues that randomly provided cash transfers from the PROGRESA experimentencourage some household members to migrate by relaxing credit constraints ofthe recipient households. Both of these studies find interesting evidence suggest-ing a role of income in relaxing credit constraints to migration in two particularcontexts. Whether this result applies widely to many situations remains aninteresting open question. In the literature on return migration, Yang (2006)exploits quasi-random variation in exchange rates caused by the Asian financialcrisis. He finds that reduced returns to working in the foreign country result inmore return migration.

Finally, because we relate wages to attrition from the public payroll, theliterature on employee retention is pertinent to our study. Models of on-the-job search (Burdett and Mortensen; 1998) rely on a margin-volume tradeoffwhereby offering higher wages results in attracting and retaining more workers.Our results can also be interpreted as evidence for such a tradeoff.

5 Identification Strategy

5.1 Main Identification

Finding exogenous variation in wages is important for a study of migration andhome wages because the correlation between wages and migration can rarely beinterpreted as the causal impact of wages. Individuals with high ability generallyreceive higher wages and migrate more frequently (Hanson, 2008). As a result,the correlation across individuals between home wages and migration will notreflect the causal impact of wages on migration. Meanwhile, the correlationbetween migration and wages across different locations will also not generallyreflect the causal impact of wages because causality also runs the other direction:migration is a supply shock potentially affecting wages.

The wage reforms described above help alleviate these difficulties. The scenedepicted in Figure 5 closely mimics an experiment with variation in the intensityof treatment. Conditional on time fixed effects, dummies for each gradeXstepcombination, and no promotions, the vast majority of the variation in wagesoccurs as a result of the policy-induced wage changes from 2005 to 2006. Dueto their sudden appearance in an otherwise stable environment, we take the

7

variation in these wage changes across different gradeXstep combinations asplausibly exogenous. As such, we estimate the impact of wages on attritionfrom the public payroll using a difference-in-differences estimator:

Ait = α+ βwit + δ′xit + gradeXstepi + ηt + εit (2)

where Ait is an indicator of attrition from the payroll; wit represents wagespaid in the public sector in Ghana; ηt is a time fixed-effect, gradeXstepi is adummy for which grade-step group an individual is in when they first enter thedata; and xit is a vector of controls for individual demographics, labor marketconditions abroad, and other domestic policies. We will estimate equation (4)by instrumental variables, using wages that an individual would have receivedif never promoted as an instrument for actual wages. Formally, if log wages aredefined as:

wit = ft (gradeit, stepit)

where changes in ft (·) embody policy changes to the public health sector wageschedule. Then we define an instrument wit as:

wit = ft (gradei0, stepi0)

where gradei0 and stepi0 denote the grade and step of individual i when wefirst observe them on the public payroll. Importantly, this instrument eliminatesvariation in wages that comes from promotions because this variation is likely toreflect ability and thus be endogenous. When combined with time fixed effectsand group effects, the variation remaining is that caused by enactment of thenew wage schedule.

5.2 Potential Confounding Factors

The sudden reforms of 2006 and otherwise stable wage environment create a rea-sonable natural experiment in which to measure the impact of wages. However,this identification strategy relies on the assumption that wage changes from 2005to 2006 for different groups of workers can be taken as exogenous. We controlfor time and group effects, which remove the influence of time-variant factorsthat affect all groups of workers similarly and time-invariant differences acrossdifferent groups. However, our identification could fail if the government ofGhana directed the largest wage increases to groups of workers based on char-acteristics correlated with attrition. As such we control for the demographicand job characteristics available in our data including age, gender, district ofjob placement, and department of job placement.

Aside from the wage reforms, other similarly-timed domestic policies alsolikely affected migration. In Ghana, several measures were taken with the goalof reducing migration of health workers. The Ghana College of Physicians andSurgeons opened in 2004, becoming the first medical specialist training schoolin Ghana. It is thought to have decreased migration of doctors who would haveotherwise migrated for training purposes. Also during this time period, the

8

Ministry of Health and the Nurses and Midwives Council collaborated with theservice delivery agencies to enforce a bonding scheme for nurses. Under thisprogram, publicly-trained nurses were required to complete a term of publicservice or pay a bond before they could be given verification by NMC to prac-tice abroad. Other policies to reduce migration included a subsidized car loanscheme as well as increased availability of fellowships for continuing professionaleducation.

Outside of Ghana, major policy changes also occurred, particularly in theUK. In 1999, the UK National Health Service adopted a Code of Practice lim-iting recruitment from developing countries. This policy strengthened consid-erably in subsequent years as the UK moved to limit not just the NHS, butalso recruitment agencies working on behalf of the NHS. Meanwhile, wages anddomestic supply of health workers in the UK were changing and could alsocoincide with the wage reforms that we are studying in Ghana. For both for-eign and domestic policies, if they affect all health workers equally, then ouridentification will still be valid. However, if any policy disproportionally affectsparticular groups of health workers, then this could bias our results. Figure 6demonstrates that this does not appear to be a problem with regard to changesin UK demand for foreign health workers. Overall inflows of health workersfrom the world into the UK have decreased dramatically, but the change hasaffected doctors and nurses similarly. In any case, we will demonstrate thatour identification strategy is not affected by controlling for the impacts of bothforeign and domestic policy changes.

Of course, unobservable factors could affect the validity of our identificationstrategy. For example, if wage increases are targeted at groups whose attritionrates are already decreasing, this will bias the estimates downward. This is animportant issue that we address by controlling for group-specific time trends.Another possible concern relates to ability of the workers. While controllingfor group effects at a fine level and removing promotions from our instrumentshould reduce bias generated by ability and positive selection, unobserved abilitycould still generate endogeneity. For example, our specification focuses on thecontemporaneous effect of wages on attrition, but if individuals make migrationdecisions based on lifetime expected wages, then our instrument will systemat-ically underestimate home wages for high ability individuals who expect to bepromoted. Since migration likely selects on ability, this would induce correlationbetween our instrument and the error term. We expect that the magnitude ofbias created by this type of endogeneity will be small because of how finely ourgroups are defined; however, we cannot rule out such a possibility.

6 Data

6.1 Administrative Wage Data

The main data source used in this study is individual-level payroll data obtainedfrom the Controller and Accountant General’s Directorate of the Government

9

of Ghana. This data contains payroll records from 2003 to 2009 for each healthworker classified under MOH paid by the central government including employ-ees of the Ministry of Health, Ghana Health Service, Christian Health Associa-tion of Ghana (CHAG), the teaching hospitals, and MOH training institutions.This panel data provides individual-level observations over several years, andyields a rich picture of the health labor market in Ghana. Additionally, foreach individual this data provides detailed information on employee grade (i.e.Senior Medical Officer, Chief Lab Technologist, etc.) and salary step for eachindividual. Information on age, gender (equals 1 if male), department (CHAG,GHS, Headquarters, etc.), and district of posting are also available.

We use health sector public wage schedules to map grade and step into abasic salary for each worker. As described above, a major part of the 2005-2006 salary changes hinged on the Additional Duty Hours Allowance. Thus,it is important to consider not just the basic salary but also the AdditionalDuty Hours Allowance for each worker. Since ADHA was allocated accordingto a fixed formula that depends on a worker’s category and their base pay, thepayroll data along with data on ADHA hours allotments from GHS allow us toestimate their ADHA earnings. From 2003-2005 the formula for total wages is:

wit = log(Eit + Eit

160 × 1.75× hgit)

where Eit is the basic salary of individual i at time t and hgit is the number ofADHA hours allotted to a worker in grade git, which is the grade of worker i attime t. From 2006-2009, we simply use the log of basic salary. Table 2 providessample statistics. The first column describes potential migrants (defined below),which is our main sample. Real (measured at 2004) monthly wages for this groupaverage 447 Ghana Cedis2.

As described above, our instrument for log wages will be log wages thatwould be earned by worker i if he or she was never promoted. In particular,the instrument for a worker i at time t takes the value from the time t wageschedule for a worker with the grade gradei0 and step stepi0 where gradei0 andstepi0 are the grade and step of worker i when he or she first enters the data.One complication with this definition is that from 2005 to 2006, the entire wageschedule changed from the Ghana Universal Salary Structure (GUSS) to theHealth Salary Structure (HSS). As a result, the number of steps for a givengrade changed in some instances. In these cases, the new step is approximatedto reflect the same percentage progress up the grade. Formally, starting in 2006an individual’s ‘first-observed step’ within the HSS is defined as:

stepHSSi0 = stepHSSlowest +(stepGUSSi0 − stepGUSSlowest

)× totalstepsHSS

i

totalstepsGUSSi

where stepGUSSi0 is an individual’s ‘first-observed’ step under the GUSS system;steplowest is the entry-level step for individual i’s grade and totalsteps refers tothe total number of steps in i’s grade under a particular salary schedule.

The payroll data is monthly but due to technical challenges does not coverall months prior to 2006. As a result, it is not possible to study attrition over

2About 500 USD.

10

monthly intervals. Thus, we analyze the data at intervals approximating oneyear. To this end, we only use data from selected months: November 2003, July2004, December 2005, December 2006, December 2007, December 2008, andJuly 2009. Since we use time fixed effects, the varying lengths of time betweenobservations should not be an important issue.

6.2 Measuring Attrition

Individual records can be matched from year to year based on identifiers in thedata. In particular we match records on first name, last name, gender, and dateof birth3. We use these matched records over time to measure attrition of healthworkers from the payroll. In particular, if an individual is in the sample at timet but never after time t, we say the individual attrited at time t. In this study,we usually interpret the impact of wages on attrition from the payroll as theimpact of wages on migration. The two are not, of course, generally equivalent.Workers could potentially leave the public payroll for employers outside thedataset (i.e. the private sector or military hospitals), or they could retire fromworking in the health field. We emphasize migration, but we cannot explicitlyseparate migration from other forms of attrition in the data.

6.3 Choosing the Sample

Interpretation of the results depends heavily on what part of the sample we usebecause the data covers workers of all ages and occupations with widely vary-ing skill sets, education requirements, and responsibilities. For older workers,retirement is a major consideration. In our data, the probability of remainingafter 20-40 years of experience is still positive and decreasing. This indicatesthat a significant minority of health workers still remains and attrite from thepublic sector after long careers. Retirement is an obvious explanation. To focuson migration rather than retirement, we truncate our main sample to those 40years of age and younger, though we will expand this scope when we investigatethe impact of wages on workers on different ages.

We also split our sample based on whether workers’ occupations allow themto be a ‘potential migrants.’ Classification as such was determined after conver-sations with administrators in the Ministry of Health and Ghana Health Service.The first column of Table 3 provides a list of all worker categories classified aspotential migrants along with their prevalence in the 2003-2006 sample. As ex-pected, nurses and physicians compose much of the ‘potential migrant’ sample,though there are also many medical assistants (nurse practitioners) and skilledallied health workers, such as laboratory and x-ray officers. The second columnindicates those excluded from this group, mostly low-skilled workers (orderlies,

3Employee numbers are available in the data and in most circumstances would be ideal.However, the treatment of some cadres’ employee numbers changed from 2005 to 2006. Giventhis re-definition of employee numbers comes at the same time as the natural experiment, wechoose to use a consistent method of matching over all time periods. Since these demographicidentifiers are nearly always unique, this method is preferred.

11

watchmen, drivers, etc.), as well as clinical workers whose skills are in low de-mand in developed countries (midwives, community health nurses, etc.). Nursesin training are available in the database but excluded from both groups becausethey are in school and receive only small stipends. The second column of Table2 shows that, as expected, the non-potential migrants have an attrition rateabout half that of the potential migrants, and they are paid less than half asmuch. However, mean age (conditional on being younger than 40) and genderdistribution are similar to our main sample: average age is around 32 and justover half of individuals are female.

Splitting the sample in this way provides two services. First, it again focusesanalysis on categories of workers that were known to have high rates of migrationduring the sample period. This helps focus the analysis on a group where thereare external reasons for believing that migration is a main cause of attrition.Second, it also provides the opportunity to test whether wages affect these twogroups in similar or different ways. Since retirement and the private sector areopen to both the potential migrants in our main sample and to those excludedfrom our main sample, if we observe similar effects of the wage changes on bothgroups, then this would indicate that we are observing the effect of wages onsome other form of attrition rather than migration. However, if the wage reformsshow no effect on attrition of non-migrants, then we are likely measuring theimpact of wages on migration.

Aside from limiting the sample in reasonable ways, circumstantial evidencealso indicates that the large scale attrition seen in the data can most plausiblybe explained by migration. As Table 2 shows, in our data roughly 8 percentof all potential migrant health workers leave the public payroll each period.As we have already seen, this coincides with well-known large-scale migrationof health workers from Ghana. So, migration is at least consistent with theattrition rates in our data. Additionally, other forms of exit seem implausible.Large-scale retirement for young workers seems unlikely. Due to central paymentof not only GHS but also CHAG and others, our data cover the vast majorityof all health workers in Ghana. Ghana MOH estimates that 81.9 percent ofall healthworkers work in the institutions covered by our sample, with most ofthe remainder in the private for-profit sector or prison and military hospitals.Also, for many categories of potential migrant health workers, coverage of ourdata frequently surpasses 90 percent (Ministry of Health (2007)). Given thesmall proportion of health workers in for-profit and military medicine, thesesectors’ human resource usage would have to increase by about 40 percent peryear to absorb all of the health workers leaving the public payroll. What dataexists on the private sector is not consistent with this story. For example, theWorld Health Organization’s National Health Accounts indicate that the privatesector’s share in health sector spending actually fell from 58.6 percent to 48.4percent from 2000 to 2007 (World Health Organization (2010)). Thus, externalevidence seems to indicate that attrition from the public payroll in our sampleis more plausibly attributed to migration than to other causes.

12

6.4 Additional Data Sources

The payroll data is supplemented by data from other sources. For inflation weuse the GDP deflator from the World Bank’s World Development Indicators. Allvalues are reported in real 2004 Ghana Cedis. In 2004, the average exchange ratewas USD 1.12 per Ghana Cedi. Foreign wages are drawn from the UK AnnualSurvey of Hours and Earnings. Individuals are matched to UK wages basedon their cadre (doctor, professional nurse, etc.) at the SIC 4-digit level, whenpossible. Others are matched to SIC 3-digit and 2-digit codes when necessary.Necessarily, these data are not nearly as precisely measured as home wagesbecause they are not individual specific. UK wages are changed into GhanaCedis using PPP exchange rates from the Penn World Tables. As Table 1shows, wages in the UK average 1121 Ghana Cedis per month at PPP rates.This is 2-3 times the purchasing power of the average domestic salary.

In addition to wages, other labor market factors in the UK likely affectedmigration during this time period, particularly adoption and strengthening ofthe NHS Code of Practice. We control for this using a measure of the opennessof the UK to migrants from different professions: the log of the total numberof new migrants to the UK for a particular profession in a given year fromall source countries. We have this variable available from registration data inthe UK with physicians from the UK General Medical Council, nurses fromthe UK Nurses and Midwives Council, and others (Art Therapists, BiomedicalScientists, Clinical Scientists, Dieticians, Occupational Therapists, Orthopists,Physiotherapists, Radiographers, and Audiologists) from the UK Health Pro-fessions Council. In our data, we also match Ghanaian dentists as physicians inthe UK; medical assistants and anesthetist assistants as nurses; lab assistantsand lab technical officers as clinical scientists; nutrition officers as dieticians;and x-ray officers as radiographers. For other categories, we code the variableas zero.

Two concurrent domestic policies affecting health worker migration are in-cluded in some specifications. The Ghana College of Physicians and Surgeonsopened in 2004, becoming the first specialist training school in Ghana. It isthought to have decreased migration of doctors who would have otherwise mi-grated for training purposes. We measure this using enrollment in the GhanaCollege of Physicians and Surgeons, with data drawn from the Ghana College’snewsletter as well as media reports. We include enrollment in the Ghana Collegeinteracted with a dummy for non-specialist doctors as a control. Also, duringthe sample period the Nurses and Midwives Council, in cooperation with GHS,began enforcing a public-sector service requirement for nurses. They beganwithholding certification from nurses who wished to migrate until they servedtheir bond period. We model this as a dummy that is one for nurses starting in2006 and zero otherwise.

13

7 Results

7.1 Main Effects

We estimate the instrumental variables regression of equation (2) with wit as aninstrument for wit. Table 4 shows the results for this regression. Coefficients onwages are normalized so that they can be interpreted as the percentage pointchange in migration resulting from a 10 percent wage increase. Each columnrepresents a different specification. First stage F-statistics are reported at thebottom of each column and indicate that a weak first stage should not be aproblem. Column (1) of Table 4 shows the results for the ‘potential migrants’sample with no covariates other than the time effects and gradeXstep effects.The coefficient is negative and statistically significant at the 1 percent level,indicating that the ‘push’ effect of wages outweighs any credit constraint effect.The coefficient of -1.59 indicates that a 10 percent wage increase would lead tojust over one and a half percentage point decrease in attrition. This is fairlysubstantial relative to the average attrition rate of about 8 percent. Column (2)adds a full set of controls for UK wages, total migration to the UK by profession,the two concurrent domestic policies, gender, a quartic polynomial in age, a setof dummies for department of posting, a set of dummies for district of posting,and gradeXstep specific time trends. Controlling for these observables, thepoint estimate remains negative but decreases in magnitude while the standarderrors increase. As a results, this effect loses statistical significance. So far, thisprovides suggestive but not conclusive evidence that changes in wages ratherthan other factors drive the results.

Contrast these results with the final two columns, which estimate the sameequations on the sample of occupations that do not tend to migrate. With andwithout controls, the effect is positive, statistically insignificant, much smallerin magnitude, and precisely estimated. Thus, we have some suggestive evidencethat higher wages decrease the probability of attrition for professions that domigrate but no evidence that such an effect exists for professions that do nottend to migrate. Since other interpretations of attrition such as retirement andmoving to the private sector are available to both groups, it seems unlikelythat these drive the results. Migration appears to be more consistent with theevidence.

7.2 Age Heterogeneity

Cutting the sample into 5-year increments by age provides additional insight.To this end, we estimate equation (2) on sub-samples of individuals age 20-25,25-30, etc. We include (previously excluded) individuals up to 65 years of ageto get a wider view on the impact across different ages. Table 5 displays the co-efficient on log wages for these different age groups, with columns denoting thedifferent specifications described previously. For the main sample, the negativeeffect is clearly strongest among the young workers under age 35, though thesmall sample of 20-25 year-olds makes that estimate imprecise. Importantly,

14

the effects for 25-35 year-olds remain both statistically and economically sig-nificant when including the full set of controls. While our results for the mainsample were only suggestive, these results provide more definitive evidence thathigher wages do in fact decrease attrition for the youngest workers. Finally, theeffects on individuals age 35-55 show coefficients that are smaller in magnitudeand statistically insignificant. The 55-65 year-old groups provide an interestingresult, suggesting a sizable negative effect of wages on the retirement decision.

As before, the last two columns of Table 5 refer to those workers on pay-roll who are unlikely to migrate. Column (3) indicates that, for this sample,our inability to detect a relationship between wages and attrition is consistentacross all age groups. The estimates are small in magnitude and statisticallyinsignificant. When we add in the controls in column (4) the results remainsimilarly insignificant with one exception. Among the older workers 50-60, wefind evidence of higher wages cutting attrition, likely due to postponed retire-ment. Altogether, these results tell a clear story. Higher wages for older workerscut attrition in both samples, but higher wages for younger workers only cutattrition among potential migrants. For older workers, higher wages can affecta reason for attrition that is common among all workers: retirement. But sincethe impact on younger workers rests only with one part of the sample, we inferthat higher wages affect a form of attrition particular to this sub-sample: mi-gration. Altogether, we take the results as evidence that higher wages decreasemigration among young potential migrants while having no effect on mid-careerworkers and decreasing retirement for all late-career workers.

The effect among early and mid-career workers is consistent with the simplestmodel of migration that focuses on how increasing home wages reduces a pusheffect. It is perhaps not surprising that these effects diminish for older cohorts.In a life cycle model of migration, a fixed cost of migration finances the opportu-nity to receive higher annual wages. With fewer years remaining, older workersface a similar cost but lower benefits to migrating. Additionally, if individualsdiffer idiosyncratically in their preference for migration, the health workers thathave chosen to remain in Ghana for several years may be a selected group thathas such a low probability of migrating that marginal wage changes have noimpact. Finally, ageing itself may be correlated with other factors (marriage,having children, building a home, etc.) that may make individuals very unlikelymigrants, again causing wages to have no measurable impact on this very smallprobability. These explanations are all observationally equivalent in our data,but they provide ample explanation for the fact that the effect diminishes inolder cohorts. In any case, it appears that these factors outweigh the impactof credit constraints, for in a model of credit constraints we would expect theyoungest individuals to have low wealth and thus be credit constrained. Then,it would be these individuals for whom the ’push’ effect would be most offsetby a more relaxed credit constraint. Clearly, we do not observe this effect asdominant in this data.

15

8 Robustness

8.1 Political Economy and Average Causal Response

In instrumental variables analysis, if treatment effects are heterogeneous acrossindividuals then the estimated coefficient is a local area treatment effect (LATE)for those individuals who comply with treatment. In our context of variable in-tensity of treatment, the analogous concept is Average Causal Response (ACR).ACR measures a weighted average of LATE’s, where the LATE’s are expecta-tions over the compliers who are moved above a particular treatment level bythe instrument and weights give weight to treatment levels in proportion withthe number of compliers (Angrist and Pischke (2009)). For our purposes, wageincreases may be uncorrelated (given covariates) with the error term, but theymay still be directed toward groups with high or low values for β, creating anACR that may not be representative of the impact of wage increases on a ran-dom sample of potential migrants in Ghana. For example, the government maydirect wage increases toward individuals with the lowest (most negative) β’s toreduce migration by the largest amount. To the extent that in future actionthe government will direct wage increases in a similar manner, then the effectmeasured here is actually more useful than the impact on a random sample ofhealth workers. However, exploring this more closely will help with interpretingthe results.

In an ideal world we would test for heterogeneity of β by extending themodel for this case explicitly to estimate βgs specific to each group:

Ait = α+ βgswit + δ′xit + gradeXstepi + ηt + εit (3)

Then, we could estimate βgs and test for the impact of βgs on wage increases:

w2006gs − w2005

gs = α+ γ ∗ βgs + δ ∗X2003gs + εgs (4)

However, due to the fact that our instrument only varies across groups andtime, estimation of the first equation is not identified. We consider the followingequation instead:

Ait = α+ βgswit + δ′xit + gradeXstepi + εit (5)

where there is no time fixed effect and for simplicity, we use the reduced formof regressing directly on the instrument. Without the time fixed effect, theequation is identified, but the estimates of βgs will be biased due to correlationwith the error term. Suppose there are no covariates (including foreign wages),then we can express the true model in equation (5) in differences from groupmeans:

∆Ait = βgs∆wit + ∆ηt + ∆εit

where the ∆ indicates difference from gradeXstep mean. Then, if we estimateβgs without the time fixed effect, we can express the bias in this manner:

16

βgs = Cov(∆Ait,∆wit)V ar(wit)

= βgs + Cov(∆ηt,∆wit)V ar(wit)

Importantly, if the correlation between wages and the time-effect component ofthe error is constant across groups, then the bias is likewise the same for allgroups. This is exactly what occurred during the pre-2006 period. All groupsreceived common percentage raises and thus have the same correlation with thecommon time fixed-effect. Thus:

βgs = βgs + ψ

from 2003-2005. This common bias is absorbed by the constant in equation (4),giving us a consistent estimate of γ, i.e. the extent to which wage increases aretargeted at groups based upon their responsiveness to wages.

We estimate the βgs parameters from two different models, with βshortgs cor-

responding to a equation (5) with no covariates and βlonggs including all controlsexcept group-specific time trends, which we exclude due to only having two timeperiods before 2006. Table 6 provides the final results of regressing 2005-2006log wage changes on the estimated βgs parameters. Columns (1) and (2) findnegative coefficients of -0.0045 and -0.0051, both of which are statistically signif-icant, indicating that wage increases may have been aimed toward groups thatwere more responsive to wage increases (more negative βgs). When we includecontrols for average age and proportion of females in columns (3) and (4), theseresults remain very similar.

These results suggest that the government directed raises at those groups ofworkers who were most responsive to such increases. Thus, our results likelymeasure a treatment effect of greater magnitude than what would be measuredif wages were randomly assigned. In this sense, our results are an upper boundfor the effect of wages on attrition. However, it seems likely that future payincreases will also be directed toward those categories of workers who are mostresponsive to salary changes. In this context, we measure an average causalresponse that is relevant to policymakers who are choosing not only the levelbut also the distribution of pay increases.

8.2 Sample Selection and Hazard Models

Due to the nature of the data available, the sample available at any time periodt is necessarily the sample of those who have chosen to remain. This selectedsample is not random, and it is well-known that selected samples can createbiased estimation if the unobservables determining selection are correlated withthe regressors (Heckman 1979, Vella 1998). If our instrument is uncorrelatedwith these unobservables, then this source of bias should not be present. Givenour natural experiment approach to instrument selection, there is reason tobe confident that sample selection bias does not affect our results significantly.However, one way to check for the robustness of this assumption is to directlyestimate hazard models. The dependent variable of interest in our study is theprobability of leaving the sample conditional on remaining in the sample untilthat time, i.e. a hazard rate. There is a large class of hazard and survival

17

models designed for this type of dependent variable that explicitly model thetendency of individuals to leave the sample (or more traditionally, die).

One type of model explicitly models the hazard itself. We adopt the com-monly used Cox Proportional Hazards model, where the hazard is assumed tobe:

λ (τ |zit; θ) = λ0 (τ) exp (β0 + β1w∗it − β2wit + δ′xit + gradeXstepi + ηt)

where zit is the data, θ is all of the parameters, and τ is the length of timesince an individual began working for the Ghanaian government. Importantly,the duration dependence, λ0 (τ), that allows the probability of migrating varyover the course of a career does not need to be parameterized. Only the impactof the covariates needs to be parameterized (here as an exponential). Giventhis assumed hazard form, a quasi-likelihood function can be formed so that theparameters can be estimated consistently without estimating λ0 (τ). Also, weuse the instrument w rather than actual wages in an attempt to preserve ournatural experiment. Given strong correlation of the instrument with wages, thisapproximation should be quite good.

We estimate this model by quasi-maximum likelihood and display the resultsin Table 7. The coefficient on the wage instrument is negative and statisticallysignificant for both models.4 The coefficients themselves have little meaning;as such, hazard ratios are often reported. Here we report the ratio of thehazard rate (probability of migrating) with a 10 percent wage increase relativeto the hazard rate without the increase. The resulting hazard ratio is 0.81with no controls and 0.70 with the full set of controls, indicating that a tenpercent wage increase would decrease the hazard of attrition by 19 to 30 percent.This is actually slightly stronger than our previous result, where a 10 percentwage increase decreased a mean 8 percentage point migration rate by 0.9 to 1.6percentage points, i.e. by 10 to 20 percent.

9 Conclusion

This paper measures the impact of home wages on attrition of skilled healthprofessionals from the public sector in Ghana by exploiting variation in wagescaused by a policy-induced natural experiment. We find that a 10 percent wageincrease reduces the annual attrition rate by 0.9 to 1.6 percentage points. Thiseffect is concentrated solely among young workers age 20-35 who come fromprofessions that tend to migrate. We take this as evidence that the effect ofwages on attrition from the public sector mainly runs through migration. Unlikeprevious attempts at measuring the impact of home wages on migration throughmacro data or parametric selection models, we find this negative effect of homewages on migration to be economically significant and robust to specification.These results support the most basic economic models of migration in whichindividuals choose to migrate based on wage differentials between home and

4Caution should be taken in inference, as the standard errors cannot be clustered

18

foreign countries. These results run counter to expectations that the impactof marginal wage changes may be small when wage gaps are very large or thathigher home wages might relax credit constraints causing more migration.

As with all experiments, natural or designed, high internal validity does notguarantee external validity of the results. Context is important. The individualsin this study are highly-skilled health professionals who tend to migrate perma-nently with their families. The results will not necessarily apply to low-skilledor low income individuals who may face tighter credit constraints. However, thesample studied here is both under-studied and of policy significance. While ouruse of attrition from public payroll rather than an explicit measure of migrationcreates uncertainty about whether we are measuring other forms of attrition,it does allow us to approach the phenomenon of permanent migration on theindividual level with a broad sample. In doing so, we find evidence for a vitalelement of migration theory in a sample characterized by long-term, permanentmigration.

The migration behavior of skilled health workers is of particular policy im-portance. Health policy and strengthening health systems in particular havegained great notoriety recently. In this context, migration of skilled health pro-fessionals away from developing countries has been widely debated. To many, itis a notorious ‘brain drain’ of needed health workers from places where skilledhealth professionals are already scarce (e.g. Chen and Boufford (2005)). As aresult, many paths have been taken toward reducing such migration. For exam-ple, the UK National Health Service’s has voluntarily imposed restrictions onforeign recruitment via its Code of Conduct, and this idea has been taken uprecently by the World Health Assembly. To others, restrictions on migration areseen as violating human rights of migrants, ignoring more important problems ofhealth worker performance and urban-rural distribution, or driven by recoup-ing misguided education subsidies (e.g. Clemens (2009)). While important,deciding this debate is certainly out of reach for this article.

However, we do demonstrate that health workers can be retained, not justby restrictions on leaving but also by rewards for staying. If policymakers indeveloping countries desire to retain more health workers, as most do, then ourresults indicate that increasing salaries is one effective option. If policymakersin developed countries desire to redistribute health workers, subsidizing healthworker salaries in sending countries is one means to this end that does not in-volve restrictions on the movement of individuals. Our results also indicate thatmigration of health workers should be an important consideration as policymak-ers in developing countries contemplate public sector wage reforms. As Ghanaconsiders transitioning all public workers to a Single Spine Salary Structure, fu-ture policy research needs to more closely examine how such a change in publicsector compensation would affect migration of health workers.

Of course, salaries for health workers are expensive, and other factors enterthe cost-benefit calculus. For example, when dealing with fairly rigid publicpay schedules, raising wages for health workers can sometimes lead to calls forwage increases in other sectors of the public payroll. While a complete cost-benefit analysis is beyond the scope of this paper, we do make progress toward

19

measuring the cost-effectiveness of using higher salaries to reduce attrition ofskilled health workers. This opens the way to compare salary increases withother policy options.

20

References

[1] Angelucci, M. (2004) ‘Aid and Migration: An Analysis of the Impact ofProgresa on the Timing and Size of Labour Migration.’ IZA Discussion PaperNo. 1187.

[2] Angrist, J. and J.S. Pischke (2009) Mostly Harmless Econometrics: An Em-piricist’s Companion. Princeton University Press, Princeton.

[3] Bhargava, A. and F. Docquier. (2007) ‘A New Panel Data Set on Physicians’Emigration Rates (1991-2004).’

[4] Bhargava, A. and F. Docquier. (2008) ‘HIV prevalence and migration ofhealthcare staff in Africa.’ World Bank Economic Review, 22.

[5] Borjas, G. (1987) ’Self-Selection and the Earnings of Immigrants’ AmericanEconomic Review, 77(4).

[6] Burdett, K. and D. Mortensen. (1998) ‘Wage Differentials, Employer Size,and Unemployment.’ International Economic Review, 39(2).

[7] Chen, L. and J.I. Boufford (2005) ‘Fatal Flows: Doctors on the Move’ NewEngland Journal of Medicine, 353(17).

[8] Clark, X., Hatton, T. J., and Williamson, J. G. (2007) ‘Explaining US im-migration, 1971-1998.’ Review of Economics and Statistics, 28.

[9] Clemens, M. (2007) ‘Do Visas Kill: Health Effects of African Health Profes-sional Emigration.’ Center for Global Development Working Paper 114.

[10] Clemens, M. (2009) ‘Skill Flow: A Fundamental Reconsideration of Skilled-Worker Mobility and Development.’ Center for Global Development WorkingPaper 180.

[11] Dovlo, D. and F. Nyonator (1999) ‘Migration of Graduates of the Uni-versity of Ghana Medical School: A Preliminary Rapid Appraisal.’ HumanResources for Health Development Journal, 3(1).

[12] Gibson, J. and D. McKenzie (2010) ‘The Microeconomic Determinants ofEmigration and Return Migration of the Best and Brightest: Evidence fromthe Pacific.’ Journal of Development Economics, forthcoming.

[13] Hanson, G. (2008) ‘International Migration and Development.’ Commissionon Growth and Development, Working Paper No. 42.

[14] Heckman, J. (1979) ‘Sample Selection Bias as a Specification Error’ Econo-metrica, 47.

[15] Lopez, R. and M. Schiff (1998) ‘Migration and the Skill Composition of theLabour Force: The Impact of Trade Liberalization in LDCs.’ The CanadianJournal of Economics, 31(2).

21

[16] Mayda, A.M. (2008) ”International migration: A panel data analysis of thedeterminants of bilateral flows” Journal of Population Economics, forthcom-ing.

[17] Ministry of Health (2007) ‘Human Resource Strategies and Policies for theHealth Sector: 2007-2011.’

[18] Okeke, E. (2009) ‘An Empirical Investigation of Why Doctors Migrate andWomen Fail to Go For Screening.’ Unpublished dissertation, University ofMichigan.

[19] Pedersen, P. J., M. Pytlikova and N. Smith (2004) ‘Selection or networkeffects? Migration flows into 27 OECD countries, 1990-2000.’ IZA DiscussionPaper 1104.

[20] Vella, F. (1998) ‘Estimating Models with Sample Selection Bias: A Survey’Journal of Human Resources, 33(1).

[21] Vujicic, M., et. al. (2004) ‘The Role of Wages in the Migration of HealthCare Professionals From Developing Countries’ Human Resources for Health,2(3).

[22] World Health Organization (2010) World Health Statistics 2010 WHOPress, Geneva.

[23] Yang, D. (2006) ‘Why Do Migrants Return to Poor Countries? Evidencefrom Philippine Migrants’ Responses to Exchange Rate Shocks’ Review ofEconomics and Statistics, 88(4).

[24] Yang, D. and H. Choi (2007) ‘Are Remittances Insurance: Evidence fromRainfall Shocks in the Philippines.’ World Bank Economic Review, 21.

22

Table 1: Destinations of Migrant Health Workers from Ghana

Destination Nurses Physicians

UK 71% 56%US 22% 35%South Africa - 6%Canada 3% 1%Australia 2% -Other 2% 2%

Source: Ghana Nurses and Midwives Council Dovlo and Nyonator (1999)

Table 2: Sample Statistics

Sample Potential Migrants OthersAttrition 0.08 0.036

(0.27) (0.19)Real Ghana Wage 447 179

(239) (93)Real UK Wage 1133 582

(774) (315)Age 32 32

(4.3) (4.9)Gender 0.48 0.42

(0.5) (0.49)Nursebonding 0.29 0

(0.46) (0)COPS 22.2 0

(46.2) (0)Log UK Migrants 6.7 0.1

(3.7) (0.9)N 22,117 54,900

Source: Administrative payroll data. Standard deviations are in parentheses.The sample includes only individuals under age 40.

23

Table 3: Defining Potential Migrants

Potential Migrants Others

Worker Category Number of Observations Worker Category Number of Observations(2003-2006) (2003-2006)

PROFESSIONAL NURSE 25,761 ORDERLIES 18,138MEDICAL OFFICER 3,648 COMMUNITY HEALTH NURSE 16,149MEDICAL ASSISTANTS 1,773 HEALTH/WARD ASSISTANT 11,941TECHNICAL OFFICER 940 MIDWIVES 8,724TECHNICAL OFFICER (LAB) 858 ENROLLED NURSES 7,582SPECIALIST 805 WATCHMEN 3,413MEDICAL OFFICER - HOUSEMAN 790 ACCOUNTS OFFICERS 3,057BIOMEDICAL SCIENTIST 669 DRIVERS 2,836HEALTH SERVICE ADMINISTRATOR 538 ARTISANS 2,733ANESTHETIST ASSISTANT 511 PHARMACY TECHNICIANS 2,420TECHNICAL OFFICER (XRAY) 488 BIOSTATISTICS ASSISTANTS 2,174TECHNICAL OFFICER (NUTRITION) 482 FIELD TECHNICIANS 2,154FACILITY/DISTRICT DIRECTOR 134 LABOURERS 1,940HEALTH RESEARCH OFFICER 117 TYPISTS 1,887HEALTH EDUCATORS 109 ACCOUNTANTS 1,697PHYSIOTHERAPISTS 106 DISPENSING ASSISTANTS 1,595TECHNICAL OFFICER (ORTHOPEDIC) 102 KITCHEN ASSISTANTS 1,588PHYSIOTHERAPY ASSISTANT 98 SCAVENGERS 1,588DENTAL SURGEON 96 TECHNICAL OFFICER (CDC) 1,511MINISTRY-LEVEL DIRECTOR 94 STOREKEEPERS 1,417OCCUP. THERAPY ASSISTANT 89 CONSERVANCY LABOURERS 1,407RADIOGRAPHERS 75 MEDICAL RECORD ASSIST 1,396HEALTH PLANNERS 64 TECHNICAL OFFICER (BIOSTATS) 1,315OPTOMETRIST 58 PHARMACISTS 1,099DENTAL TECHNOLOGISTS 49 SECURITY GUARDS 1,014NURSING TUTOR 30 LABORATORY ASSISTANTS 899HEALTH EDUCATION OFFICER 28 WASHMEN/IRONERS 868OCCUPATIONAL THERAPISTS 4 STENOGRAPHERS 856

24

Table 4: Main Effects

(1) (2) (3) (4)

Sample Potential Migrants Potential Migrants Others OthersDependent Variable Attrition Attrition Attrition Attrition

Log Ghana Wage -1.59*** -0.89 0.12 0.33(0.60) (0.91) (0.16) (0.36)

Log UK Wage – -0.14 – -0.06(0.16) (0.05)

Log UK Migrants – -0.0028 – 0.000(0.0034) (0.001)

Nursebonding – 0.12*** – –(0.03)

COPS – 0.00037** – –(0.00017)

Gender – -0.030*** NO 0.012***(0.005) (0.003)

Age Quartic NO YES YES YES

Time Fixed Effects YES YES YES YES

GradeXStep Effects YES YES NO YES

Department Dummies NO YES NO YES

District Dummmies NO YES NO YES

GradeXStep Specific Time Trends NO YES NO YES

N 21,378 21,347 51,729 51,543Number of Clusters 329 327 971 967First Stage F 69 45 2561 400

Statistical significance at the 1, 5, and 10 percent levels is denoted by ***, **, and * respectively.Standard errors are clustered at the gradeXstep level.

25

Table 5: Effects by Age Group

(1) (2) (3) (4)

Sample Potential Migrants Potential Migrants Others Others

20-25 -6.20*** 7.20 -0.30 -1.19(1.28) (20.12) (0.62) (2.87)

25-30 -1.77*** -1.78** 0.01 0.29(-0.60) (0.72) (0.32) (0.76)

30-35 -2.58*** -2.27** 0.34 0.26(0.55) (1.11) (0.31) (0.61)

35-40 0.66 1.72 0.14 0.57(0.73) (1.10) (0.24) (0.57)

40-45 -0.39 -0.63 0.00 -0.05(0.80) (1.05) (0.23) (0.51)

45-50 -0.10 -0.81 -0.00 -0.25(0.53) (0.66) (0.15) (0.31)

50-55 -0.74 -0.59 -0.13 -0.74**(0.69) (0.76) (0.15) (0.35)

55-60 0.50 -2.83** 0.04 -1.06**(1.74) (1.38) (0.59) (0.53)

60-65 -0.80 -2.78 3.47 6.79(5.85) (5.62) (3.00) (6.00)

Controls NO YES NO YES

Statistical significance at the 1, 5, and 10 percent levels is denoted by ***, **, and * respectively.Standard errors are clustered at the gradeXstep levelCoefficients of a regression of attrition on log Ghana wage reported.

26

Table 6: Testing ACR

Dependent Variable: ∆ Log Wages ∆ Log Wages ∆ Log Wages ∆ Log Wages(1) (2) (3) (4)

βshortgs -0.0045*** – -0.0056*** –

(0.0013) (0.0014)

βlonggs – -0.0051*** – -0.0057***

(0.0011) (0.0011)Mean Age 2003 – – 0.0007** 0.0006**

(0.0003) (0.0003)Proportion Female 2003 – – 0.008* 0.006

(0.005) (0.004)N 104 162 104 162R-Squared 0.113 0.129 0.161

Statistical significance at the 1, 5, and 10 percent levels is denoted by ***, **, and * respectively.

27

Table 7: Cox PH Model

(1) (2)

Hazard Ratio 0.81 0.70

Instrument for Log Ghana Wage -21.0*** -35.1***(4.5) (8.1)

Log UK Wage – -6.5***(1.4)

Log UK Migrants – -0.04(0.04)

Nursebonding – 2.0***(0.4)

COPS – 0.0039(0.0031)

Gender – -0.30***(0.08)

Age Quartic NO YES

Time Fixed Effects YES YES

GradeXStep Effects YES YES

Department Dummies NO YES

District Dummmies NO YES

GradeXStep Specific Time Trends NO YES

N 14,685 14,669

Statistical significance at the 1, 5, and 10 percent levels is denoted by ***, **, and * respectively.Results are from a Cox-PH survival model; standard errors are not clustered

28

Figure 1: Historical Migration of Physicians from GhanaSource: Bhargava and Docquier (2007); the data have been converted into flows

Figure 2: Migration and Attrition of Nurses from GhanaSources: Nurses and Midwives Council; IPPD Payroll Database

29

Figure 3: Migration and Attrition of Physicians from GhanaSources: UK General Medical Council; IPPD Payroll Database

Figure 4: Salaries of Health Workers in Ghana and the UKSource: IPPD Database; UK Annual Survey of Hours and Earnings; UK figures are PPP

30

Figure 5: Wages for Health Workers in Ghana: 2003-2008Each line indicates the real log wages of a particular gradeXstep group (e.g. senior medical officers onstep 5); each group’s wages are normalized to zero in 2003; Source: IPPD Database

31

Figure 6: Migration of Health Workers to the UK from All Source CountriesSources: UK General Medical Council and UK Nurses and Midwives Council

32